Chinese Researchers Unveil MemOS: A Groundbreaking Memory Operating System for AI Models

Researchers from Shanghai Jiao Tong University, Zhejiang University, and other leading institutions in China have unveiled MemOS: Memory OS for AI System, the first operating system designed specifically for managing the memory of AI models. MemOS aims to address a fundamental limitation that hinders AI models from emulating human brain functions more closely and enhances their capacity to learn from experience.

This new system, MemOS, treats memory as a crucial computational resource, managing it similarly to how processors and storage are handled in traditional operating systems. A study published on July 4th in arXiv demonstrates significant performance improvements compared to current approaches, including a 159% increase in temporal reasoning tasks when compared to OpenAI’s memory systems.

Modern AI faces the challenge of «isolated memory,» an architectural limitation that prevents them from maintaining coherent, long-term relationships with users. Each session starts from scratch: models cannot remember user preferences, accumulated knowledge, or behavioral patterns from previous interactions. This diminishes the user experience; for example, an AI might forget about a user’s dietary restrictions mentioned earlier when asked for restaurant recommendations again. This issue is particularly pronounced in corporate settings where AI is required to retain context over complex processes that unfold over weeks.

While some solutions, such as retrieval-augmented generation (RAG), attempt to address this by pulling in external information during a session, researchers argue that such methods remain ineffective patchwork solutions. The core issue is not merely information retrieval, but rather the creation of systems capable of truly learning and evolving based on experience, akin to human memory.

MemOS proposes a fundamentally different solution using what are known as “MemCubes” — standardized memory blocks that can hold various types of information, migrate, merge, and evolve. These encompass both textual knowledge and adaptations of parameters and activation states within the model, forming a unified memory management framework that has not existed before.

In tests using the LOCOMO benchmark, which evaluates memory-intensive tasks, MemOS consistently outperformed all existing approaches. The system achieved a 38.98% improvement over OpenAI’s memory implementation, particularly excelling in complex logical reasoning tasks where it had to connect information from different parts of a dialogue.

Additionally, the system demonstrated up to a 94% reduction in latency (time until the first token) thanks to a new KV-cache memory integration mechanism.

These achievements indicate that memory, previously considered a bottleneck, is a more significant constraint than previously thought. By treating memory as a computational resource, MemOS adds reasoning capabilities that were previously «constrained» by existing architectures.

The architecture of MemOS is inspired by classic operating systems, structured around three levels: the interface level (API), operational level (scheduling and memory management), and infrastructure level (storage and access management).

The core component, MemScheduler, dynamically manages both temporary and persistent memory, selecting storage and retrieval strategies based on different tasks and usage patterns. This starkly contrasts with current methods, where memory is either static (as determined by model parameters) or ephemeral (limited to the current context).

The emphasis shifts from how much a model knows to whether it can convert experience into structured memory and reproduce it consistently. This is a prerequisite for an architectural shift in AI—from pre-training to experience-based learning.

The team has made the MemOS code available on GitHub, with plans for integration into Hugging Face, OpenAI, and Ollama. Currently, MemOS supports Linux, with plans to extend to Windows and macOS.

Thank you for your attention, from your Cloud4Y team. Follow us here or on our Telegram channel!